Dario Amodei published a policy essay in June 2026 arguing that the United States should stop relying on transparency rules and start regulating frontier AI like the Federal Aviation Administration regulates airplanes. Alongside the policy essay, Anthropic released a legislative proposal on frontier model testing and a framework for job displacement, both of which the company says it will financially back.
What he's actually asking for
The core demand is mandatory third-party testing. Models above a compute threshold would have to be evaluated for risk in four areas: cybersecurity, biological weapons, loss of control, and automated R&D that could speed up the other three. And here is the part that separates this from everything Anthropic asked for before. The government would get the power to block or reverse a model's release if assessors decide it presents unacceptable risk in those four categories.
Amodei frames the airplane comparison carefully. Cars, planes, drugs: powerful, essential, and capable of killing people if built badly. Testing before release, not lawsuits after harm. He's explicit that this could later escalate if the most capable systems start looking less like aircraft and more like "weaponizable nuclear materials," his words, but he doesn't want to legislate for that scenario yet.
The reversal
This is a notable shift, and Amodei doesn't hide it. Back in 2024 he sent two public letters about California's SB 1047 expressing, in his own description, mixed feelings. He worried about exactly the failure mode he now wants to avoid: rigid legislation that burns compliance effort on requirements that turn out not to matter while missing the risks nobody anticipated. His August 2024 letter to Governor Newsom said the amended bill's benefits likely outweighed its costs, then immediately hedged that he wasn't certain and parts of it seemed ambiguous. That is not a full endorsement, and it wasn't meant to read as one.
Through 2025 Anthropic narrowed its asks to disclosure. The company backed transparency laws including California's SB 53 and pushed a federal transparency standard. The logic was that you collect evidence first, then write precise rules once the risks have a clear shape.
So what changed?
Amodei points to one thing: Claude Mythos Preview. He credits it with proving frontier models pose real cybersecurity threats to financial systems, critical infrastructure, and national security, and says it moved AI models into the category of tools with strategic consequence. He expects biological risks to follow, then autonomy risks. Whether you take that at face value depends on how much you trust a company to be the sole narrator of why its own product proves regulation is now urgent. The evidence cited traces back to Anthropic's internal red-teaming, which is worth keeping in mind.
The timing also isn't happening in a vacuum. On June 2, 2026, the Trump administration signed an executive order setting up a voluntary review process, where developers can give the government early access to models for up to 30 days before release. An earlier draft reportedly allowed 90 days before it got cut. Amodei calls the order an incremental move in the right direction, then says his proposal goes further. That gap is the whole point. Voluntary and 30 days versus mandatory with a binding off-switch are not small differences in degree.
The part that's hard to take seriously
The reasonable-sounding pieces of the essay, the testing regime, the labor framework, the macroeconomic support, all assume a state apparatus willing and equipped to act on them. Amodei half-acknowledges this himself with a Treebeard metaphor about institutions that move too slowly to matter. He thinks the forest is finally waking up. The harder question, which the essay mostly waves at rather than answers, is who on the government side is positioned to receive the sensible half of these proposals and actually build them.
Anthropic's legislative proposal and job displacement framework are out now. The Trump executive order's directives carry implementation timelines of 30 to 60 days, so the contrast between what the administration is building and what Amodei is asking for should sharpen over the next two months.
